# E72-J/K — Verified memory and governed self-improvement

## Memory (E72-J)

20 typed engineering facts with provenance, written to AETHER-33-backed collective memory; 20 **held-out**
questions scored by exact expected content (`artifacts/e72/memory-eval.json`).

| Arm | Correct |
|---|---:|
| No memory (answers UNKNOWN) | 0/20 (0.00; Wilson 0.00–0.16) |
| Verified-memory retrieval (top-3) | **20/20 (1.00; Wilson 0.84–1.00)** |

Retrieval precision@3 = 0.433; p50 latency 1.68 ms. Supersession returns the current verified fact and hides the
older one by default. An injection-like write is **QUARANTINED** and never recalled. Contradictions are flagged, not
auto-resolved. Every hit carries `grants_authority:false`; 0 authority violations.

The no-memory arm is deterministic (it answers UNKNOWN), so the 0→20 delta measures **retrieval availability**, not
reasoning. A real local model arm is smaller but more informative: **0/6 without context → 4/6 with retrieved
context** (SmolLM2-360M, exact substring scoring; two answers were semantically right but failed the strict token).

## Self-improvement (E72-K)

Two experiments were proposed, measured on held-out work, and left **AWAITING_OWNER_APPROVAL**
(`artifacts/e72/self-improvement.json`):

| Target | Baseline | Candidate | Delta |
|---|---:|---:|---:|
| model routing (P1 → P2) | 0.250 | 0.500 | **+0.250** |
| scheduling admission (no controller → adaptive) | 0.333 | 1.000 | **+0.667** |

The registry refused a protected governance target, a missing held-out set, evaluator substitution, and
**agent self-promotion**. History hash chain verifies. Nothing was enabled in the running system.